{"id":"arxiv/2602.05657","url":"https://arxiv.org/abs/2602.05657","tags":["cs.LG","math.OC"],"score":0,"title":"Tight Long-Term Tail Decay of (Clipped) SGD in Non-Convex Optimization","domain":"cs.LG","source":"arxiv","authors":["Aleksandar Armacki","Dragana Bajović","Dušan Jakovetić","Soummya Kar","Ali H. Sayed"],"reviews":0,"subject":"cs.LG","summary":"The study of tail behaviour of SGD-induced processes has been attracting a lot of interest, due to offering strong guarantees with respect to individual runs of an algorithm. While many works provide high-probability guarantees, quantifying the error rate for a fixed probability threshold, there is a lack of work directly studying the probability of failure, i.e., quantifying the tail decay rate for a fixed error threshold. Moreover, existing results are of finite-time nature, limiting their ability to capture the true long-term tail decay which is more informative for modern learning…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-02-05","reactions":{}}